Field Notes
The first arc was orientation.
This one is inspection.
Once you see where this is going, you start looking at software differently.
Before, the question was simple:
Does this tool help my team?
That question still matters. But it is no longer enough.
The better question now is:
Can my agents use this without turning me back into the operator?
That sounds like a small shift. It is not.
It changes how you evaluate your entire stack.
A tool can be beautiful and still be useless in the new world. It can have a clean dashboard, a modern interface, great branding, and a sales team that says AI in every other sentence. But if the only way to use it is for a human to log in, click seven buttons, copy a value, paste it somewhere else, and remember what happened, that tool is not agent-ready.
It is a locked room.
And the street outside is getting automated.
I have become much more aggressive about this in my own systems. If a tool cannot be read by an agent, operated by an agent, or audited after an agent touches it, I no longer see it as neutral. I see it as future drag.
That does not mean every tool needs to become an autonomous robot. It means the tool needs a door.
The door might be an API.
It might be a command line interface.
It might be MCP.
It might be a webhook.
It might be clean exports, clean logs, or a browser surface that can be safely operated with permissions and review gates.
But there has to be some way for software to participate in work without forcing a person to become the bridge every time.
This is why I switched from ManyChat to Boosend for my Instagram comment-to-DM flows.
Most people know the front-end experience. You comment a keyword under a reel, and an automated DM shows up with the guide, link, or signup page.
That part is useful.
The operator side matters.
With ManyChat, I still had to log in and build each automation. Pick the post. Set the keyword. Configure the message. Connect the next step. Check the flow. It worked, but I was still the bridge.
With Boosend, the workflow matches how I actually think:
“For my next reel, when someone comments New World, send them my AI newsletter signup and the free AI guide.”
That sentence is the test. The value is not the DM. The value is that the system can build that workflow from intent.
Done.
The agent can set up the process. No logging in. No clicking around. No rebuilding the same automation by hand.
That is the difference.
The future does not belong to the company with the most software.
It belongs to the company whose software can be used by the most capable workers, human and AI, without everything routing through one exhausted person with a browser tab open.
Most companies still buy tools the old way: features, interface, adoption, integrations.
All useful. But the new world adds a sharper question:
Could an agent complete a real workflow here and prove what it did?
If the answer is no, the tool may still be useful today. But you should understand what it is costing you tomorrow.
It is turning your people into middleware.
It is making your operators manually carry information between systems.
It is forcing your best people to stay trapped in manual loops because the software was designed for a world where only humans did the work.
That world is ending.
Playbook
Here is the first-pass audit I would run on any important tool in your business.
Pick one tool your company depends on.
Not a random app. Pick something that actually matters. Your CRM. Your project management system. Your accounting tool. Your inbox. Your scheduling tool. The place where work lives.
Then ask these questions.
1. Can an agent safely authenticate into it?
This does not mean sharing your personal password with a bot.
The standard is safer than that.
Can you create a separate account, token, key, or controlled access path? Can permissions be limited? Can access be revoked?
If the only access path is “use Brian’s login and hope nothing weird happens,” the tool is not ready.
2. Can an agent read the current state?
This is the first real test.
Can the system expose what is true right now?
Open tasks. Customer records. Calendar events. Messages. Orders. Files. Payments. Tickets. Whatever the tool manages.
If an agent cannot read the state, it cannot reason about the work. It can only guess, and guessing is not operations.
Look for API docs, exports, webhooks, searchable history, data views, or some clean way to retrieve the facts.
3. Can an agent take action?
Reading is only half the job.
Can the agent create the task, update the record, send the draft, move the card, schedule the meeting, or trigger the workflow?
Not every action should be automatic. In fact, most important actions should start with approval gates.
But the tool should at least make action possible.
If a human must do every final click forever, the tool is not a work surface for agents. It is a waiting room for humans.
4. Can the system stop at the right places?
Agent-ready does not mean reckless.
The best systems have clear stop signs.
Draft the email, but do not send it.
Prepare the pricing recommendation, but do not push it live.
Create the invoice draft, but do not charge the card.
Summarize the document, but do not delete or move files.
Good agentic systems are powerful because they are bounded.
5. Can it leave receipts?
This is the part most people skip.
If an agent touches a system, you need to know what happened.
What did it read?
What did it change?
What did it recommend?
What did it send?
What failed?
What still needs human review?
Without receipts, you do not have an operating system. You have vibes with access.
Receipts can look like logs, sent items, comments, version history, saved reports, task updates, audit trails, screenshots, or structured run notes. The format matters less than the principle.
The work needs to be inspectable.
6. Can it complete one real workflow?
Do not audit in theory.
Pick one workflow and walk it through.
For example:
Can an agent read a customer request, find the account context, draft a response, create a follow-up task, and leave a summary?
Can an agent read a weekly operations mess, turn it into priorities, draft team messages, and show what needs approval?
If the answer is yes, you have an agent-ready surface.
If the answer is “only if a human keeps copying things between tabs,” you found the bottleneck.
7. Mark the tool
You do not need a complicated scoring system.
Use four labels.
Keep: agents can read, act, and leave receipts with reasonable permissions.
Wrap: the tool is useful, but it needs a bridge, script, export, API helper, or operating procedure around it.
Replace: the tool traps important work behind human clicks and has no serious path to agent access.
Watch: the tool is not agent-ready yet, but it is not critical enough to replace today.
This is how the stack starts changing.
Not through a giant transformation deck.
Through one practical question repeated over and over:
Can my agents actually use this?
Orientation
This is why I think a lot of companies are about to misread the AI shift.
They will buy chat features and think they modernized.
They will add an assistant button to a dashboard and think they are agentic.
They will let employees paste screenshots into AI tools while the underlying company stack remains closed, manual, and unauditable.
That is cosplay.
That is a chatbot sitting outside a locked building.
The real shift is deeper.
Work is moving from human-only interfaces to mixed human-agent operating systems. The winners will have better surfaces for work to move through.
The people who understand this early have an advantage inside their companies.
You can walk into the room and ask the question everyone else is missing:
Are we buying tools for humans to click, or are we building a stack agents can operate?
That question turns AI from a novelty into an operating lens.
Next, we need to define what an agent actually is. Because if we are going to audit tools this way, we need to separate a chatbot from a system that can read, reason, act, stop, and leave proof.
That is the next post.
For now, start with one tool.
Comment below with one piece of software your business depends on, and whether an agent could actually use it.
— Brian